Using process mining and event log analysis for better business strategy decision-making
Yuri van Midden
University of Twente P.O. Box 217, 7500AE Enschede
The Netherlands
j.h.vanmidden@student.utwente.nl
ABSTRACT
Process mining is often used for conformance checking and performance checking in businesses. However, after es- tablishing the problems, seeking solutions and deploying them, that is often the end of the process, whilst event logs analysis can also be used to yield business perfor- mance. This paper proposes methods to utilise process mining and event logs analysis to aid business decision makers in narrowing down the considered business op- tions, and yield business performance. After defining how event logs analysis can be used to execute performance checking and thus narrowing down business options, a case study is conducted wherein the methods are tested. Three experiments are conducted in which at the start of the event logs all possible scenarios are considered, and after a given interval the worst performing scenarios are elimi- nated. The experiments show that the proposed methods of narrowing down business scenarios all converge to simi- lar best-performing scenarios, and that event logs analysis has great potential to aid business decision makers.
Keywords
process mining, event log analysis, business strategy, lo- gistics, simulations
1. INTRODUCTION
For business decision makers, it can be hard to make strate- gical decisions that improve the business and often there are time and money constraints that restrict testing and evaluating all possible business scenarios. As resources, time and money are limited, there is a call to make deci- sions that promise most business performance.
Process mining is an emerging discipline relating to data science and BPM discovery, aiming to support in the field of business process analysis and management[5]. Often process mining is used to do performance and conformance checking in a business, to answer questions like: “Why does it take so long to get a product from A to B?” or
“What part of a process takes a lot of time?”
Even though process mining provides answers to a lot of these questions, process mining is often used a as goal in itself, instead of as a means to improve business decision- Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy oth- erwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee.
35
thTwente Student Conference on IT July. 2
nd, 2021, Enschede, The Netherlands.
Copyright 2021 , University of Twente, Faculty of Electrical Engineer- ing, Mathematics and Computer Science.
making and strategy. Furthermore, process mining and event logs analysis are scarcely used as tools for business decision-makers to evaluate different scenarios whilst the business is operating. Usually event logs are considered as a whole, instead of narrowed down and evaluated to smaller time spans.
Hence the aim of this research is to propose a method to help business decision-makers in determining a strategy for effectively narrowing down the number of considered busi- ness scenarios. The narrowing down of business scenarios should be time efficient and substantiated by statistical arguments. We propose an approach to steer in select- ing scenarios that are expected to yield the most promis- ing business performance. This is achieved by evaluating event logs of different scenarios over given time spans, af- ter which - at given intervals - a scenario-selection is made to continue for evaluation.
To achieve this goal, the following research questions are answered:
1. What are suitable performance indicators in a busi- ness in the logistical domain?
2. How can event logs be used to derive quality mea- sures?
3. How can event log analysis be used to narrow down possible business scenarios to comply with the re- quirements of the performance indicators?
The answers to the research questions will be applied to a case study as a means of validation.
The remainder of this document is structured as follows:
first some background information on event logs and pro- cess mining will be provided. Thereafter the research ques- tions will be answered, followed up by the case study that integrates the previously accumulated findings. Lastly, the results will be presented and a conclusion will be drawn.
2. BACKGROUND 2.1 Event logs
Events in the context of event logs and process mining are actions that are recorded in a log. They typically con- sist of data like the start time, completion time, activity description, allocated resources, cost and a case ID [6].
For accurate process mining it is important that events
always consist of a case ID and a timestamp, in order to
be able to do performance and conformance checking. A
brief example of an event log can be seen in table 1. An
ingredient, milk, comes into a factory and is being moved
around a forklift vehicle. The sequence of such actions
combine into an event log. Lastly, event logs do not need
Table 1. Example event log for a factory processing fresh ingredients, with quality measures
timestamp ingredientID
and quality event resource
9:00:00 milk1:100% enter factory Forklift A 9:01:25 milk1:98% picked up Forklift B 9:03:02 milk1:93% dropped off Forklift B 9:03:03 milk1:93% start processing Machine A
... ... ... ...
Figure 1. An example Petri net, adopted from [7] (page 27)
to origin from a single source, they can be a combination of multiple prepared data sets.
2.2 Process Mining
With the event logs at hand, process mining can be used for process discovery, conformance checking and perfor- mance analysis [6]. With process mining, it is possible to identify event traces from event logs, and create a model from the discovered traces. An example of such a mined process model (i.e. Petri net) can be found in figure 1.
Process mining is intended to answer several questions in the context of a given business. Typically these questions are divided into two categories: performance questions and conformance questions [7]. Performance questions may be “Why are these products always late?” or “What re- sources are overloaded?” whilst conformance questions can be “Which part in the process is often skipped?” or “What business rules are often violated?” Using the event logs and the discovered Petri net model, an event log replay on the Petri net can be executed to monitor the performance of the events on the given traces. This also enables for evaluating what strategies result in business rules being violated.
We determine the following perspectives: conformance check- ing perspective, time & case perspective, organizational per- spective and the data perspective.
2.2.1 Conformance checking
Conformance checking is divided info four core principles, according to Van der Aalst: fitness, simplicity, precision and generalization [5].
We can determine the fitness of an event trace. A process model has a perfect fitness if all traces in the log can be replayed by the model from beginning to end [5]. We can define fitness by looking at the case (for example, milk1 from the example in table 1), and determine whether all events in the case exist in the process model from the beginning to the end. Another way of determining fitness is by checking if an event is represented in the process model at all. If all these analysis turn out to be true, then the fitness is 1. If none of the cases are represented by the model, the fitness is 0.
Second the simplicity of a model can be checked. Van der Aalst [7] describes the method of Occam’s Razor, which comes down to that the simplest representation of the be- haviour of the event logs is probably the best one.
Third, precision is related to fitness. The model is precise when it does not allow for more types of behaviour then can be imposed by the (potential) events.
Last, generalization is opposite to precision. When a model is too specific, it may fit for some specific event logs, but not be able to fit other event logs. Therefore a model should be generalized enough.
For businesses in the logistics domain, we can determine several threshold values that trigger conformance moni- toring. For instance, we often hear ”a 99% success rate”, or ”99.9% success rate”, which can be translated into con- formance checking threshold values. In this example the fitness of the event traces should always be above 99 or 99.9, and if the threshold is exceeded (i.e. value drops below 99), the business scenario may be altered to yield better performance again.
2.2.2 Case & time perspective
Second the case-and-time perspective is related to perfor- mance checking in process mining. Executing performance analysis in tools like ProM [8], aids in determining the time for cases to get from the beginning to the end in the model, as well as some basic statistics like the mean, me- dian, standard deviation, minimum and maximal time for some or all cases.
For businesses in the logistics domain, performance indi- cators such as the mean of the clearing times of all cases or the mean + standard deviation of the clearing times of all cases.
2.2.3 Organizational perspective
Thirdly, key performance indicators (KPIs) are consid- ered, as they represent the means of the company in order to achieve their goals. For example, if a product at the end of the production line must have a ”quality value” of at least 90%, it can be argued that this value should be a threshold for elimination business scenario options.
2.2.4 Data perspective
Lastly, the data perspective is considered, which is most important for quality measures in this case. The event logs often contain a lot of hidden data that can be used for performance analysis, one being the case & time indicator.
An other factor one can imagine is whether customers need to go through a lot of repetitive processes, making the amount of repetitions a performance indicator. Lastly, it is possible to infer data from the event logs, and use this data for performance measures. For example, if a cost attribute is included in the event logs, one can classify the costs as high, medium or low, which can be evaluated in the process discovery phase as separate processes. This classifier method can aid in enriching the existing data sets.
3. RESEARCH QUESTIONS
This research aims to describe methods to narrow down the number of considered business scenario options, using process mining and event log analysis. These methods will thereafter be applied to a case study. First, the research questions mentioned in the introduction will be answered in this section:
1. What are suitable performance indicators in a busi-
ness in the logistical domain?
2. How can event logs be used to derive quality mea- sures?
3. How can event log analysis be used to narrow down possible business scenarios to comply with the re- quirements of the performance indicators?
3.1 Performance indicators
This research will mainly focus on processing event logs in a logistical context, therefore we need to consider possible actors and their interests in the business. Furthermore we will analyse what Key Performance Indicators are essential for the operations. These will form a basis upon which we can carry out our case study.
There are several performance indicators that can be con- sidered in a logistical context. For example, a CEO wants to keep the waiting times for a delivery as low as possible, while maintaining low costs, and by using as little person- nel as possible. For answering this research question, we want to determine what the indicators are that make us switch to a different strategy if needed.
Let us consider a factory that processes food. The factory uses several fresh ingredients as an input and produces a final product that is shipped to grocery stores. The input ingredients are likely to show a quality decay over time and therefore we define a few performance indicators, such as ingredient quality at the start, processing time, amount of personnel and vehicles needed to transport the ingredients through the plant. Furthermore the quality of the end product should be high enough and the transport to the grocery store should go as smooth as possible. The key performance indicators in this example can be formulated as follows:
• The time from raw ingredients to finished product does not exceed 3 hours.
• The temperature of the ingredient throughout the process should not exceed 20 degrees centigrade.
• The waiting time from finished product to transport to the grocery store should not exceed 2 days.
• The quality decline of the ingredients should not ex- ceed 20%
If we take a look at the event logs a factory described above may generate, we can imagine a few data points that are included in the event logs, like product quality, the timestamp, the unique identifier of the ingredient, the resource that is handling the ingredient. An example can be found in table 1.
It is not hard to see that some of these data points in this table are suitable for quality and performance check- ing. For instance, the time it took for a forklift vehicle to transport the goods can be a quality measure, or the quality decay of the ingredients themselves.
Krauth et al.[3] have created a framework that indicate suitable performance indicators for logistical service providers.
They divided their framework in internal and external KPIs, where internal means that the KPIs are relevant for the company’s management and employees, and exter- nal means the customer and society as a whole. Internal KPIs for managerial use are for example number of deliv- eries, trips per period, average fuel use per km, % of failed orders and human resource costs.
3.1.1 Process mining and event log analysis
Using process mining and a continuous analysis of the event logs, the aforementioned KPI’s success rate can be monitored and evaluated. Even more, when a business scenario does not yield much performance and the KPI cannot be met, event logs analysis can help uncover this and eliminate such scenarios from the set of considered options.
3.2 Using event logs for quality measures
Event logs are capable of holding other data fields than just the case ID, timestamp and action, such as a temper- ature measure, or a quality attribute. As Mannhardt et al. [4] state, these attributes are often not used in pro- cess discovery, and therefore result in unreliable quality diagnostics for the discovered models. Therefore, meth- ods are sought to perform analysis based on these quality measures.
One method described before, is the time & case perfor- mance indicator where individual event traces are classi- fied based on the case times. A method to classify these case times is by using the mean case time and the stan- dard deviation to determine the human-readable score of the cases. An example can be found in table 2. These per- formance classes can be combined with the events in the event logs, which enables for performance checking using process mining (i.e. how often does a trace go from ‘good’
to ‘insufficient’ ?).
Table 2. Basic classifier for performance values value performance class
< µ − 2σ very bad
< µ − σ bad
< µ insufficient
< µ + σ sufficient
< µ + 2σ good
> µ + 2σ very good
3.3 Narrowing down business scenarios
In this section methods are proposed to narrow down busi- ness scenarios for decision-makers, using event logs analy- sis. Possible scenarios one can think of are combinations of resource allocation (personnel and vehicles), investments and costs, KPIs, machinery, and so forth. First, the life- cycle of a typical data mining project will be discussed using the CRISP-DM model [9].
3.3.1 CRISP-DM
Let us consider the CRISP data mining model, in which
one considers the life cycle of a data mining project in a
given business. The model is portrayed in figure 2. Sum-
marised, the model tells us how a data mining project
should be assessed: we typically start with Business Un-
derstanding, where we aim to figure out what business
intelligence we need in order to solve a problem. From
there we proceed to Data Understanding, where we collect
data and get familiar with the available data. Next, we
process the data in such a way that we can get a final and
usable data set from the raw data, the Data Preparation
phase. Next we start the Modelling phase, which essen-
tially means we tweak the data sets and calibrate it to
optimal values, in order to ensure we can tackle the data
mining project. In the Evaluation phase we ensure that
the obtained model is exactly what is need to tackle the
data mining problem, and we come to a final decision in
whether we deploy any solution in the business. As can
Figure 2. ”Phases of the CRISP-DM Process Model for Data Mining.” Derived from [9]
be seen in the figure (2), some phases can be executed in any desired order.
3.3.2 Methods for narrowing down business sce- narios
In order to determine the best performing business sce- narios, several methods can be applied when analysing event logs. The methods effectively touch upon every as- pect of the CRISP-DM model. In order to define certain goals, business understanding and data understanding are needed. In order to define our strategy for narrowing down business scenarios, accurate data preparation and mod- elling are needed.
The aim is to define methods that narrow down business scenarios, so that the most promising business scenarios remain. For example, let us consider a set of event logs (n=100) of a factory of different periods in time with alter- ing business scenarios. One is able to play through these event logs simultaneously and evaluate the conformance and performance after every day, or every hour, or even at any given moment in time. The conformance and perfor- mance can be determined using process mining and basic statistics, and thus more promising business scenarios can be derived.
For this research, various event log replay strategies have been defined. For example: “Play through the event logs simultaneously and after one hour, ...”
• “...discard half of all the scenarios that performed worst. Then repeat with the remaining logs.”
• “...discard the single worst performing scenario. Then repeat with the remaining logs.”
• “...calculate the mean score of all scenarios and dis- card every scenario that performed worst than the mean value. Then repeat with the remaining logs.”
• “...discard any scenario that has a lower score than the mean − standarddeviation. Then repeat with the remaining logs.”
• ...
These findings will be put into practise in the case study below.
4. CASE STUDY
The case study is intended to test the findings of the re- search questions. The case study is building upon previous research by Bemthuis, R. [1] where a conceptual agent- based simulation framework is proposed, to analyse and learn from emergent behaviour in complex business sit- uations. The simulations that were run involved a fac- tory that processes certain products, which were moved around by various types of vehicles: human-driven fork- lifts (HDF), automated guided vehicles (AGV) and un- manned aerial vehicles (UAV), visualised in figure 3. An overview of the factory is depicted in figure 4. We can see the factory is partitioned into region 1, region 2 and region 3. Typically, during the simulations the products were transported from region 1 to region 2 to region 3.
Certain dispatching rules were imposed on the simulations, distinguished by so-called vehicle-initiated and product- initiated rules. The vehicle-initiated rules were: random, pick lowest product quality decay, pick highest product quality decay. The product-initiated rules were: random, call lowest utilisation, call shortest travel distance. Fur- thermore simulations were run with altering compositions of the vehicles.
All permutations of the different scenarios run in the sim- ulations, resulted in data sets of 27 different scenarios on which we conduct our experiments. The scenarios can be viewed in Table 3. The scenarios are divided into three resource allocations:
• Scenarios 1-9: 3 UAVs, 1 HDF, 1 AGV
• Scenarios 10-18: 3 UAVs, 2 HDFs, 2 AGVs
• Scenarios 19-27: 2 HDFs, 2 AGVs
Table 3. Scenarios simulated in the data sets random lowest
quality
highest quality random {1,10,19} {8,17,26} {9,18,27}
lowestUtilization {2,11,20} {4,13,22} {6,15,24}
shortestTravelDist. {3,12,21} {5,14,23} {7,16,25}
The final goal of the case study is to narrow down the possible business scenarios whilst analysing the event logs, and yield the supposedly most promising scenario for a business decision maker. As a performance indicator the quality of the products when dropped of in region 3 is used.
4.1 Methodology
The data sets, obtained from [2] include event logs of all different scenarios in table 3. Using the knowledge ac- counted for in section 3.3, various experiments are run to narrow down the possible business scenarios to yield the best performing scenario. By executing the various meth- ods, the experiment converges to the best strategy after a certain period.
The raw data sets are of the format portrayed in table 4.
One can see that a product arrives at the factory in region 1, is the picked up and transported to region 2, and finally dropped off in region 3, which is considered the end point.
The experiments will focus mainly on the currentDecayLevel
as a performance measure. This value is the representa-
tion of the quality of the product throughout the process,
between 100 and 0, where 100 is high quality and 0 is low
quality.
Figure 3. Visualisation of various vehicles in the simulations, adopted from [1]
Table 4. Sample taken from event logs used in the experiments, in this case: scenario 1 (Exp1.txt)
uniqueID productIDstr event timeStamp vehicle currentDecayLevel
... ... ... ... ... ...
969 .Models.MUs.product3:15 arrivalAtSource 00:06:54.7458 - 100.0%
970 .Models.MUs.product3:15 productCallsTransportRegion1 00:06:54.7458 - 100.0%
971 .Models.MUs.product3:15 assignedToVehicleRegion1 00:06:54.7458 UAV:2 100.0%
972 .Models.MUs.product3:15 pickedUpRegion1 00:06:54.7458 UAV:2 100.0%
... ... ... ... ... ...
984 .Models.MUs.product3:15 droppedOffRegion2 00:07:00.9958 UAV:2 98.7%
985 .Models.MUs.product3:15 startProcessingRegion2 00:07:00.9958 UAV:2 98.7%
... ... ... ... ... ...
1658 .Models.MUs.product3:15 droppedOffRegion3 00:10:57.4733 UAV:4 97.1%
... ... ... ... ... ...
Figure 4. The case study ’factory plant’, adopted from [1]
4.1.1 Process discovery
To get an overview of the event traces in the event logs, process discovery is applied. The discovered process model is portrayed in figure 5. In this instance the process mining tool Disco
1was used for the process discovery, with the default settings and filtered on complete event traces (so that incomplete event traces are discarded). This process model applies for all scenarios provided in the data sets.
4.2 Experiments
For this research, several experiments are defined to nar- row down the business scenarios. Every experiment uses a short algorithm that returns a ranking of the scenarios over a given time span.
The steps to determine the performance of scenarios is summarised as follows:
1. Traverse through event logs from timestamp A to timestamp B
2. Filter on all events called droppedOffRegion3 3. Of the remaining events, calculate the mean of all
values of currentDecayLevel and save it
1
https://fluxicon.com/disco/
From now on this algorithm will be referred to as the sce- narios performance algorithm. The resulting value is con- sidered the performance measure in these experiments. A high result value means that the scenario yielded better performance than a low result. This is algorithm is ex- ecuted for every scenario, after which a ranking can be made of every scenario over the time span from A to B.
Pseudocode for this algorithm can be found in Appendix A, algorithm 1.
When running the experiments, the algorithm starts at the start of the event logs (e.g. timestamp 0 ) and traverses from time span AB, to time span BC, to time span CD, and so forth.
The different experiments that were executed for this re- search were summarised as follows:
1. After every iteration, use the top half of the scenarios according to the performance value and iterate once again until one scenario remains.
2. After every iteration, use the scenarios that scored higher than the mean of the performance values of all scenarios and iterate once again.
3. After every iteration, discard the scenario that yielded least performance and re-iterate with the other sce- narios.
All experiments were run using Python and the Pandas
2data analysis library for Python. The data sets were loaded and then processed using the described steps. The exper- iments themselves are explained in more detail below.
4.2.1 Experiment 1
In the first experiment, after every iteration of the sce- narios performance algorithm, the resulting performances are sorted from best performance to worst performance.
Then the lower half of the scenarios are discarded, and the scenarios performance algorithm is executed on the
2